WO2002067422A2

WO2002067422A2 - N dimensional, non-linear, static, adaptive, digital filter design using d scale non-uniform sampling

Info

Publication number: WO2002067422A2
Application number: PCT/US2002/000940
Authority: WO
Inventors: Philip Druck
Original assignee: Philip Druck
Priority date: 2001-01-05
Filing date: 2002-01-04
Publication date: 2002-08-29
Also published as: AU2002306410A1; EP1360761A4; WO2002067422A9; US20040139135A1; EP1360761A2; US20090244395A1; CA2437510A1; US7467172B2; WO2002067422A3

Abstract

An N dimensional signal processor using D scale non-uniform sampling is disclosed. The processor includes a multi-resolution sampler (A) for outputting a stream of non-uniformly spaced samples resulting in an overall average sampling rate sufficient to satisfy Nyquist criteria to reconstruct the signal, a filter bank router (B) for mapping each point in the output stream to a filter in a filter bank, a digital filter bank (C) where each filter in the bank does filter of a subset of the samples at different uniform rates such taht the stream is uniformly spaced, and a consolidator (D) for consolidating or interleaving uniformly spaced streams form the filters into one non uniformly spaced stream.

Description

N Dimensional, Nonlinear, Static, Adaptive, Digital Filter Design Using D Scale Non- Uniform Sampling

FIELD OF THE INVENTION

This invention concerns N dimensional digital filter design in itibroadest sense. Filters are used to restrict or sculpt ID signals, 2D/3D images or more abstract N dimensional objects that traverse communications channels They are also used for data smoothing and prediction, signal detection and noise reduction applications.

Typically, digital filters rely on the Finite Impulse Response (FIR) and Infinite Impulse Response (IIR) linear mathematical models. Data inputs to these models are almost always assumed to be uniformly spaced sample points Various filter architectures use these underlying models to accommodate diverse application categories, to include: a. Static filtering (e.g., baseband, passband, notch) b. Adaptive and equalizing filters c. Weiner or matched filtering d. Kalman filtering

This invention also concerns novel uses of digital filters to modulate positive and negative feedback loops in control systems as well as to enhance digital phase lock loops.

Further, this invention extends the domain of applicability of digital filter techniques to a certain range of nonlinear systems.

Finally, this invention encompasses the derivative work of exactly calculating the n-th level autocorrelation and cross-correlations of N dimensional signals as well as convolutions.

BACKGROUND OF THE INVENTION

This discussion is confined to one dimensional (ID) filters because the N dimensional case is readily generalized from the ID case.

A ID filter is characterized by a vector of possibly complex valued weights or parameters. These weights are the tap values in a transversal filter arrangement. These weights are derived from the discrete approximation to the continuous convolution of an input signal and the impulse response of the system to filter. The convolution characterizes the filter's impact on the input signal, as follows:

+ oo y(t) = I h(t - τ) u(τ) dτ

- 00

where u(τ) is the input signal and h(t - τ) is the (non-causal in this case) impulse response of the system shown in Figure 1.

The impulse response of a system is the result of applying a sharp spiked signal at the input. The spiked signal is a discrete approximation to the unrealizable impulse of zero duration and infinite amplitude (defined using the Dirac delta functional). The idealized spectrum of impulse is flat with unit amplitude and constant phase.

The discrete approximation of the impulse response is:

+ oo k=N-l y(t) = J h(t - τ) u(τ) dτ « 1/N Σ h(n - k) u(k) - oo k=0 where 1/N is the uniform, Nyquist, sampling rate used to digitize the continuous signal.

Digital filters for linear systems are categorized as Finite-Impulse Response (FIR) or Infinite Impulse Response (IIR) in architecture. The IIR occurs when the output at one sample time is fed back as input to the next sample event. Figures 2 and 3 depict a general purpose FIR or IIR digital filter, respectively, with both architected as a transverse filter.

There are many varied techniques for designing digital filters. Static filter design entails the determination of static tap weights for specific problem domains, bounded by requirements and resource constraints. Adaptive digital filters result when the tap weights are allowed to vary, typically depending on a comparison of the actual output to a reference output. Typically, the Weiner-Hopf equation provides the analytical basis for calculating an adaptive filter's optimal tap weights. CURRENT STATE-OF-THE-ART DESCRIPTION OF KNOWN SOLUTIONS

The current state of the art in digital filter design encompasses a wide range of solutions to the underlying equations that determine the filter weights These techniques are generally categorized by the filter architecture, to include a. Linear Filter Design Algorithms b. Somewhat Nonlinear Filter Design Algorithms c Adaptive and Equalizing Filters d. Werner or Matched Filters e. Kalman Filters

Deficiencies of Existing "Prior Art" - The prior art is generally based on the following premises a Linear systems models. b Approximation of the Fourier Transform via the Discrete Fourier Transform (DFT) This is a fair approximation under best signal condition, and potentially considerably worse for transient and non-stationary signals. c. Generally periodic, stationary signals d. Decoupling of sampling from filtering operations That is, sampling is performed with whatever approach independently of the filtering results e. Uniform sampling of the input signal

BRIEF DESCRIPTION OF THE INVENTION

The N dimensional digital filter design is generalized in a straightforward manner from the ID case. Therefore the description will focus on the ID case. An N dimensional object will then be referred to as a signal, using the ID terminology.

The motivation for this invention evolved from a realization that a critical component was excluded from the typical information flow through digital filters. This flow is illustrated in Figure 4. Inputs can typically include the current input sample, previous samples, a priori knowledge and reference data. But the A/D sampler is not directly incorporated into the actual filter. Thus, standard filter design typically begins with a vector of uniformly sampled input data.

Fundamentally, non-uniform sampling of a volatile signal can glean more information about a signal than uniform sampling at the Nyquist rate (if indeed that rate is known a-priori). Additional information might be gleaned from simultaneous pseudo-random non-uniform sampling, along with trend or volatility sampling. Relative to judicious non-uniform sampling, uniform sampling can be sub-optimal in collecting information per unit of effort expended.

This filter design invention, named the D digital filter, extracts that additional information from the non-uniform sampling of signal volatility to improve the effectiveness and performance of the filter's intended application. The performance measure or metric used depends on the application. For example, for standard noise filtering and signal smoothing applications, the D filter is designed to sharpen spectrum cutoff and to minimize ripples in the frequency spectrum, while preserving phase. Or, for adaptive applications, the D filter is designed to accurately or closely track changes from a reference condition, with fast response time. Figure 5 illustrates the information flow through a filter with sampling incorporated.

This invention is of a digital filter design comprised of up to seven (7) components, shown in Figure 6. The components actually used in any implementation will depend on the application. Typically, adaptive filtering design will use many of the design components while static filters the fewest number.

The D filter design components are:

A. D Scale Multi-Resolution Input Sampler

B. Filter Bank Router

Filters can be applied to output streams following traversal through a channel as well as to inputs. Each point in the output stream might be shifted relative to its position in the input stream. This component then maps each point to a transversal filter in the bank of filters that is associated with a particular D subscale.

C. Digital Filter Bank

Each filter in the bank implements a vector of tap weights using ?he standard FIR and IIR architectures shown in Figures 2 and 3 respectively. Each filter is associated with one D subscale^, and the number of taps is the size of the D subscale. Collectively, the filter bank then implements the D Scale underlying the sampling.

D. Consolidator

This component consolidates or interleaves the multiple, uniformly sampled D Subscale streams into one non-uniformly spaced stream.

E. Arithmetic Mappings (Networked Tables)

This optional component is used to implement the various real time calculations on the sample values.

F. Sampler Controller

This component sends control signals back to the sampler, to dynamically change the sampling behavior on the input.

G. Phase Shifter

This component modifies the phase of output samples for each D subscale (via a delay). A motivation is to modulate or control positive, negative and phase lock feedback loops.

Several illustrations are presented to highlight the scope of mix-and- match of these components, Figure 7 shows the static D filter design. And Figure 8 show§ the Adaptive D filter design, both derived from the D filter components.

Further, the D Filter components might be used to modulate the input sample streams based on the output. This can be implemented using control signals that are based on the output, and then applied to the input samplers. The applied control signals would then suppress or trigger certain D subscale samplers in various frequency ranges. The Sampler Controller module can serve that function. Figure 8 shows a D filter arrangement that implements this sampling feedback loop.

Also, it might be desirable to modify the phase of an output or an input signal. The Phase Shifter component could fulfill that purpose, as. shown in Figure 9.

A D Filter can be applied in many configurations, identically to the standard digital filters. Figure 1 1 shows several possible static and adaptive D filter configurations.

The D filter design derives directly from the capabilities of the D Scale (Patent Application No. 09/326,084), non-uniform sampling using the D Scale, the D Arithmetic (Patent Application No. 09/568,368), and on the exact Fourier transform decomposition of a signal via D Scale sampling. Collectively, they provide the rigorous mathematical modeling justification for the D filter structure and operation.

A signal structure is also introduced to aid the characterization of a channel for the construction of a D digital filter. Further, industry standard static and adaptive digital filter designs are special cases of the D filter design when only unifoπn sampling is used. For then, the sampler is simply an A/D device, the filter bank reduces to one filter and there is no routing to a filter bank.

The D filter is also an evolutionary design, for it extends and extrapolates current filter design techniques. This preserves the investment in the vast installed base of digital filters.

Highlight of the D digital design include:

1. Enhanced IIR digital filtering by increasing its stability while minimizing the number of tap weights. Stability is the main constraint to wider applications of current state-of-the-art IIR filters.

2. Sharper or finer "sculpting" of a channel's bandwidth to achieve the same filtering effect as the usual designs, but with reduced number of tap weights. Alternately, the same number of taps as in a standard filter yields heightened filtering impact.

3. Capability to retrofit existing filter designs because of its backward compatibility.

4. Evolutionary design is built on the existing framework of current digital filters. This reduces implementation risk as well as preserve the investment in the installed base..

5. Mix-and-match component approach lends itself to solving many varied problem in diverse application domains.

6. D Arithmetic self-stabilizing arithmetic further enhances the

7. stability of IIRs and of adaptive filters.

8. Efficiencies due to fine level of input sample filtering as well as feedback from output back to input sampler.

DETAILED DESCRIPTION

CONTENTS

1. Detailed Description of D Digital Filter Components

1.1 D Scale Multi-Resolution Input Sampler

1.2 Digital Filter Bank

1.3 Sample Router

1.4 Multi-Stream Consolidator

1.5 Phase Shifter

1.6 Sampling Feedback Loop

1.7 D Arithmetic Tables

2. Rationale for D Digital Filter

2.1 Decomposition of Convolution Integral

2.2 Internal Architecture of D Filter Bank

2.3 D Signal

2.4 Nonlinear Systems Addressed by D Filter Design

2.5 Method for Empirically Measuring Impulse Response of a Linear or Nonlinear System

3. Application of D Filter Design

3.1 Linear Filter Design Algorithms

3.2 Nonlinear Filter Design Algorithms

3.3 Filters Derived from Weiner-Hopf Equation

3.4 Benefits of Generalized Weiner-Hopf Equation on

Adaptive Filters

3.5 Computing Autocorrelation and Cross-Correlation

3.6 Generalized Kalman Filters

1. Detailed Description of D Digital Filter Components

1.1 D Scale Multi-Resolution Input Sampler

Patent Application No. 09/326,084 describes how the D Scale sampler can be implemented by multiple off-the-shelf, commercially available analog-to- digital (A/D) converters. Each A/D device samples uniformly at a rate determined by that D Scale's subscale resolution. An extra component is required to distribute or dispatch or fan out the original analog data stream into multiple feeds, one per D subscale sampler. Then each sampler filters out all but its subscale points. Figure 12 illustrates two possible multi-sampler designs.

1.2 Digital Filter Bank

The heart of the D Digital Filter design is a bank of ordinary, commercially available IIR or FIR digital filters, as shown in Figure 2. Each filter in the bank is associated with a subscale of the underlying D Scale used for sampling. That is, each filter independently samples the incoming data at the rate of its associated D subscale.

A 1 dimensional (ID) D digital filter is characterized by a vector of of vectors of complex valued parameters or tap weights. This vector of vectors corresponds to the individual D subscales in the underlying D Scale used for the non-uniform sampling. It is realized by a bank of the usual digital filters, with each characterized by one vector of tap weights from the vector of vectors. The weights are the coefficients of the discrete impulse response.

1.3 Sample Router

Additionally, a router component directs each new sample to a filter in the filter bank. This generalizes the usuallD digital filter, that is characterized by a vector of weights. Figure 13 illustrates the architecture. Figure 14 provides a close up of the filter bank router. This component first maps each incoming point into a D Scale point, if the point is not already a D Scale point. It then routes the mapped point to that filter in the D filter bank corresponding to the D subscale containing the mapped point.

1.4 Multi-Stream Consolidator

This component consolidates the multiple D subscale uniform sample streams into one stream consisting of interleaved samples from the multiple D subscale streams. The net result is one stream of samples flowing at a non-uniform rate.

1.5 Phase Shifter

This component changes the phase of processed D subscale point streams (or signals). It performs this task by delaying the passage of each D subscale point by a predetermined value. In effect, the phase shifter transforms a point on one subscale into a point on another subscale or even onto the same subscale, but at a different frequency. A phase shifter might be useful for applications involving positive and negative feedback loops.

1.6 Sampling Feedback Loop Controller Component

The feedback loop can selectively feedback control messages or signals to the D Scale sampler at the input stream. These control signals effectively modulate the input stream that actually passes through the channel being filtered. This mechanism adds a new input sampling strategy, in addition to random and trend sampling. Figures 6 and 9 illustrated where this component can be integrated into the D filter design.

1.7 D Arithmetic Tables

The D Arithmetic is detailed in Patent Application #09/568,368. Actual floating point measurements or sampled values are mapped to D Scale points. Arithmetic operations are performed on these points by mapping the floating point result of the arithmetic operation on each pair of operands into another D Scale point. Scaling is properly preserved as well when the D Arithmetic tables are created. A table can be a physical region in memory of a computer host. Or it can be a logical table of references or pointers to a network of physical tables distributed across multiple computers.

The D Arithmetic tables can be readily integrated into a filtering system to perform the calculations, which become iterative mappings across D Arithmetic table mappings. The D Arithmetic is portable and it does not have floating point operation overflow, underflow or divide-by-zero conditions. Further, it has a self-stabilizing property that effectively bounds round-off errors to within a tolerance value. This tolerance depends on the minimum possible distance between all adjacent points in the D Scale that underlies the D Arithmetic tables.

Therefore, use of the D Arithmetic in filter design can diminish, if not mitigate entirely, the corrosive effects of floating point calculations in filter design. 2. Rationale for D Digital Filter Bank Design

2.1 Decomposition of Convolution Integral

The motivation, indeed, the underlying mathematical justification, for this design is first presented followed by details on determining the filter's tap weights.

As mentioned previously in the background section and as shown in Figures 2 and 3, the vector of weights in an industry standard digital filter derives from the standard discrete approximation of the convolution integral, of a signal and the impulse response of a channel

These N complex coefficients are the tap weights in a transversal digital filter. The rationale for this linkage of coefficients to filter taps is that the vector of tap weights approximates the impulse response of a linear system.

However, the D Scale can be applied to calculate the exact convolution of a complex function. Consider the same linear system illustrated in Figure 1. Let the D Scale be comprised of N subscales {pi, . .. PN }

Let δ,_j denote the distance of point p_j in D subscale I to the next D Scale point.. Recall that the sampling is uniform at a rate of 1/p, for each subscale I. Therefore, for points on subscale i: t = n/p, where n= 1 , ... p, - 1

(Note that for the smallest scale, p_m,n n starts at 0. That is, the convention is that the smallest scale contains the 0 D Scale point).

Further, the convolution integral can be decomposed using Riemann summation over the D subscales, identically to the procedure used to derive an exact decomposition of the Fourier Transform in Patent Application Serial # 09/326,084. The continuous convolution integral is calculated by determining its value separately at points on each D subscales.

On each subscale P_j. the integration variable τ extends over uniform intervals 1/p_j. Therefore on each subscale: τ = k/p_j

Also, on each subscale, the non-uniform interval associated with each uniformly distributed D Scale point leads to: dτ -> δ_Jιk

+ oo y(t) = J h(t - τ) u(τ) dτ = J h(τ) u(t-τ) dτ *

- 00 Decomposition of Convolution into D Subscale Filters

k= _Pj -l y(n/p,) = Σ Σ h(k/p_j) u(n/p, - - k/Pj) δj.k j e {p,, .... p_N} k=0 n=l, ... ρ, -l

This expression is then used to calculate the continuous convolution integral across the range of interest, by assembling the calculated convolution values along points on each D subscale. Again, this procedure is identical to that described in Patent Application #09/326,084 to calculate an exact decomposition of the Fourier Transform of a transient signal. The approximation becomes more exact as more D subscales are used, which increases the resolution of the D Scale.

2.2 Internal Architecture of D Filter Bank

Note that this discrete approximation to the continuous convolution integral contains N of the usual discrete impulse response convolution series, one for each D subscale. This is the key observation to establishing the link between the coefficients of these series and the D filter bank tap weights. For consider the following slight rearrangement of the terms in this equation: Let a weight be denoted by: where the asterisk denotes co

which assumes that the tap inputs, and therefore the tap weights are all complex valued. Then:

y(n/p,) = Σ Σ w*_J>k u(n/p, - k/p_j) j e {pi, .... ρ_N} k=0

The equation can now be interpreted in the language of filter weights. For each internal summation convolves the finite duration impulse response of one filter, represented by the vector w*,,ι_< for D subscale j, with the input u(n/p,) from D subscale i, to produce an output contribution y(n/p,) for a point on D subscale i. Figure 15 illustrates the internal structure of an FIR D filter bank that realizes this equation. Note that the conventional unit delay box z^"1 has an additional subscript associated with the D subscale. zo^"1. The intent is to highlight the variable "unit" interval as it depends on the D subscale.

The D filter bank of IIR filters follows the same pattern whereby the D Scale δ_J?k values are multiplied with the original weights and a set of N conventional IIR filters are tied to a D Scale sampler and router. 2.3 D Signal 2.3.1 Motivation

The preceding equations assumed that the impulse response of a system was already known. There is a then a need to actually determine or measure the impulse response of a system (or channel). The standard theoretical approach is to discretize the Dirac functional as follows:

+ oo + oo h(t) = J h(τ) δ(t-τ) dτ * Σ h(n Δτ) δ(t - n Δτ) Δτ - oo n = - oo

The effect of any input x(t) to the linear system is then determined using the system's linearity property as:

+ oo + oo y(t) = J h(τ) x(t-τ) dτ * Σ h(n Δτ) x(t - n Δτ) Δτ

- oo n = - ∞

The approximation of the continuous Dirac impulse response is a spike of unit amplitude. Note that phase is ignored in this procedure.

The motivation for the D Signal is the set of equations in Patent Application #09/326,084 on an exact decomposition of the magnitude and phase of the Fourier transform of a signal using non-uniform D Scale sampling. The equation for the magnitude decomposition is: k= Tp * p_k -l | F(2πm T_Pi)| = Σ Σ δ_pk,„,_pk h(n/p,) e ^'2π \ > ^{nm/ (}W j e {pi, .... ρ_N} k=0

The decomposition consists of a double sum of terms with each term of the form: δ Xp ,n/pk W h(_nn//„pi) e -²" (^TP_k ^fr _pi )' nm/ (T _pk * p Y)

The inverse Fourier Transform decomposition is used to determine the shape of a signal derived from a specified frequency spectrum.

2.3.2 Relation to Impulse Function

These decomposition equations can be interpreted as a formula for measuring the impact of a channel or system on linear combinations of certain atomic or elemental signals. An elemental signal is one term in the decomposition, whether forward or inverse.

Let a D Signal be one of the terms in the decomposition above. Then any (reasonably well behaved) signal can be decomposed into a set of D signals. The D signal, then realizes a D Scale point. It is characterized by a predominant monotomc frequency at the rate of the subscale's generator. If prime numbers are used, then the dominant frequency is l/p where k is the k^lh subscale in the D Scale used. In practice, there will be a spread, quantified by a variance, in the frequency spectrum of the D Signal. This variance must be within the maximum distance between adjacent points of the D Scale, as shown in Figure

16. This insures that each D Scale point's corresponding D Signal is distinct.

Thus, if a set of signals is desired with a flat frequency spectrum in a specified range, then the atomic terms can set their middle factor h() to 1. Further, if the phase is flat, then the exponential term simplifies considerably. The net effect of these constraints is a signal with a flat frequency and phase response. Indeed, by definition, that signal would realize the impulse in the defined frequency band. The D Signal approximation of an impulse response can be tailored to the frequency domain of interest, as in the microwave or optical frequency ranges.

2.3.3 Possible Uses of D Signals

One possible use of the D Signal is as CDMA signal components. Another use is to empirically measure the impulse response of a possibly nonlinear system, discussed in more depth later.

2.4 Nonlinear Systems Addressed by D Filter Design Methodology:

Nonlinear systems exhibit a wide range of behavior. Consider that one end of the spectrum of behavior is almost linear (i.e, superposition and time invariance apply) while at the other end is extreme behavior such as expected at the rim of a black hole. The scope of nonlinear systems considered by the D filter design is bounded by the part of the spectrum to include non-time varying systems that exhibit deterministic, multi-frequency generation per monotonic or monochromatic frequency input.

The following methodology can be used to quantify such nonlinearities. First, a D Scale is created for the application domain considered, such as microwave, optical or millimeter regions. Then each unit amplitude D Signal associated with the D Scale is applied at the input. Because of the system's nonlinearities, outputs at multiple frequencies can be expected, as shown in Figure 17. These output signals will have varied amplitudes and phases. And their frequencies can be mapped to D Scale frequency points. Again, the nonlinearities are assumed to be time invariant to enable reproduceable behavior. The net result of these efforts is that the output spectrum can be expressed as follows, illustrated in Figure 18A:

Where:

Yj_j ^■♦ Component of output D Signal associated with subscale i due to an input D Signal from D subscale j..

And the following results after collecting contributions from all input D Signal frequencies, as shown in Figure 18B:

N-l N-l

Yi(ω) - ∑ Yij(ω) = ∑ Hij(ω) * Xj(ω) j=0 j=0

N = number of D subscales 0 i < N-l

Where:

Hj_j •> Transfer function for input D signal associated with subscale j and output D Signal associated with subscale i.

X_j "^ Input D signal associated with subscale j.

Yj ^■> Output D Signal associated with subscale i. Or in matrix notation,

Y = H * X Where:

H is an NxN matrix each of whose elements is a transfer function H (ω) X is a IxN vector of input D Signals

Alternately, this can be expressed in the time domain. In that case, There are many impulse response functions to consider, one for each D Signal.

Note again, that for this general case contains the usual relationships between input and outputs, when the system is linear. For then, there is only one output frequency per input D Signal and the transfer function matrix only contains diagonal elements That is,

Y,(ω) = H,,,(ω) * X,(ω)

Or as usually expressed in linear system theory:

Y(ω) = H(ω) * X(ω)

Nonlinearities might also occur such that the change in output amplitude is not proportional to input amplitude changes. In that case, the same test methodology can be applied, but varying input amplitudes on each trial. That is, after a set of output signals is collected for a range of D Signal inputs at a set amplitude, a new set of D Signal inputs is applied, but with the amplitudes incremented. This procedure is illustrated in Figure 19.

2.4 Methodology for Empirically Measuring Impulse Response of a Linear or Non-linear System

The previous discussion can be summarized through the following procedural description. A flow chart illustrating the methodology is presented in Figure 20.

2.5.1 Construct a D Scale for the problem at hand:

Decide on which subscales to include in a D Scale based on a-pπor knowledge of the application domain.

2.5.2 Generate D Signals

Generate the D Signals corresponding to the D Scale points constructed in the previous step. Just use the steady state part of the signal

2.5.3 Apply D Signals

Apply the D Signals to the channel of interest, to determine the impulse response.

2.5.4 Measure the responses for each signal application. There is a matrix of input signals. Note that in the D Scale range, nonlinearities can create multiple output frequency results per input frequency Therefore each output is the composition of multiple contributions from D Signals

2.5.5 Calculate the transfer function matrix as follows

H_1J(ω) = Y_1J(ω) / X_j(ω)

2.5.5 Calculate the impulse responses by taking the Fourier Transform of the calculated transfer functions. The D Scale decomposition should be applied, rather than the standard, coarse, DFT approximation. This effort will yield a vector of vectors of possibly complex coefficients.

3. Applications of D Filter Design

3.1 Linear Static Filter Design Algorithms

There are many well known and applied categories of linear filter design algorithms, and many implementation approaches within each category. For example, one approach is frequency domain sampling, while another approach, typically used for IIR designs, relies on complex plane mappings.

State-of-the-art digital filter design approaches typically perform two general steps. First is a determination of the digital filter architecture to use, based on the application requirements, and resource constraints. The two basic architectures are Finite-Impulse-Response (FIR) or Infinite-Impulse-Response (IIR) filters. Second, a vector of possibly complex weights or transverse filter taps is calculated.

Each existing technique is readily generalized to the D filter design. For the D filter design is a direct generalization of the standard vector of tap weights. For example, the frequency sampling approach is generalized to use non-uniform D Scale sampling of a frequency spectrum, rather than the uniformly spaced samples usually applied. The same steps are then followed for the set of uniformly sampled D subscales. This same reasoning can be applied to each current digital filtering design approach.

Standard filter design entails the determination of the vector of tap weights of a filter. The three most commonly used FIR design approaches are, along with their D filter generalizations:

1. Frequency sampling

The desired frequency response is sampled uniformly at a sufficiently high rate. Then the discrete inverse DFT of the samples is calculated.

This technique is readily generalized for the frequency sampling can be performed non-uniformly using the D Scale. Then the inverse Fourier transform can be more accurately calculated using the D Scale samples.

2. Windowing

This technique entails multiplying the desired frequency response by a "windowing" function. The purpose is to reduce the impact of the Gibbs phenomenon when approximating the impulse response using the discrete DFT.

This technique is also readily generalized. For the inverse calculation can be performed more accurately allowing for more varied windowing functions. 3. Min-Max Optimization or Least-Mean-Square Error

This technique minimizes a mean-squared error criteria by solving a derived set of linear equations. Many algorithms are in use including Levison-Durbin, Parks-McClellan and Remez exchange, to note some of the most widely used ones.

This technique is also readily generalized for D filter design for the error measure would contain extra sets of terms, each set corresponding to a D subscale. This will be discussed further in the following sections, as well as in the section that discusses the derivation of the generalized Weiner-Hopf equation. Optimizing that error measure then entails the concurrent solution of multiple sets of linear equations, with each set associated with a D subscale. An attraction is that each set of linear equations can be of an order that is possibly significantly lower than that of the standard technique. This is due to the reduced number of sample points (or sampling rate) for each D subscale compared to the far higher rate needed by the usual technique.

The most commonly used IIR filter design technique is:

1. Bilinear Transformation In this general approach, a transfer function is represented as a fraction containing poles and zeros, for the most general IIR case. The generalized D filter transfer function can also be expressed as either a series of such fractions, or combined into one fraction. Then the same standard techniques could be applied.

Several other techniques, less commonly known or used , are available to design filters using both magnitude and phase specifications.

3.2 Nonlinear Filter Design Algorithms

The same enhancements to the standard design techniques apply to the nonlinear systems described in Section 2.4. For this class of nonlinear systems can be characterized by an extension of the sum of sum of tap weight vectors described previously. Therefore the extra vectors in the nonlinear systems are accommodated by simply extending the descriptions in the previous Section 3. 3.3 Filters Derived from the Weiner-Hopf Equation 3.3.1 Conventional Weiner-Hopf Equation

The Weiner-Hopf equation underlies the design of adaptive filters as well as of recursive, predictive, Kalman filter types. It is derived from a linear model that relies on uniformly sampled input and output values, shown in Figure 21. An equation is derived by seeking the extreme point of a least-mean-square expression to yield, using the nomenclature from "Adaptive Filter Theory" by S. Haykin, 3^rd edition,

R ω = P where:

R = E[ u(n) u*(n) ]] -> MxM autocorrelation matrix of input u(n) - Mxl tap input vector [u(n) , u (u-M+1) ] ^τ

ω = [w₀, W_M-_I] ^T -^ Mxl vector of transversal filter's tap weight that optimize an error criteria.

P = E[ u(n) d*(n) ]] - Mxl cross correlation matrix of input-output

The solution is a vector of possibly complex valued tap weights in the transverse filter. This equation is solved iteratively by descending an M+l dimensional error performance surface that is shaped as a bowl. It has M degrees of freedom represented by the filter's tap weights. The optimal solution is at the bottom of the bowl, as shown in Figure 22. At each step:

(Updated value of) = (Last value of) + (Learning rate) * (Tap input vector) (tap weight ) (tap weight ) (parameter ) (error signal )

This recursively determined solution is realized by dynamically changing tap weights in an adaptive filter.

3.3.2 Generalized Weiner-Hopf Equation for Linear Systems A generalized Weiner-Hopf equation for linear systems is derived from a mathematical model that uses non-uniformly sampled input and output samples, as well as the D Signal based impulse response, described in Section XXX. The standard Weiner-Hopf equation and model is shown as a special case when uniform sampling is applied. The generalized equation forms the foundation for D filter design.

The generalized Weiner-Hopf equation is derived using the same methodology as that for the standard equation. Figure 23 presents the model used by the derivation. The derivation relies on the observation that the optimal value of a set of series of terms is located by optimizing each series of terms in the equation. In this case, the set of series occurs because each is associated with one of the D subscales. What results is a set of smaller dimensional conventional Weiner-Hopf equations. The set of such equations can be organized into a super matrix Weiner-Hopf equation to yield:

Generalized Weiner-Hopf Equation

R ω = P where:

R = E[ u_pι(n) u_PJ*(n) ]] -> Matrix of NxN autocorrelation matrices of

D subscale inputs. Each element is an autocorrelation matrix whose elements are correlations between input points on D subscale i and j, for all D subscales. N is the number of subscales in the D Scale. u_pι(n) -> M,xl tap input vector [u_pι(n) ..u_pι (u-M,+l) ] ^τ where each sample is in D subscale i. M, is the number of samples that depends on The D subscale i.

ω = [W₀, W_N-_I] ^T - Vector of Nxl vectors W,=[w_l0, w_iMl.ι] ^τ

Each vector contains the tap weights for the transversal filter in a D filter bank associated with D subscale i.

P = E[ Uj(n) d_j*(n) ]] -> Matrix of Nxl cross correlation matrices of

D subscale input-outputs. Each element is a cross correlation matrix whose elements are correlations between input and output points from D subscales i and j respectively.

i and j -> 0 <= ij < N where N is the number of

D subscales.

The solution is a vector of vectors of possibly complex valued tap weights in each D filter bank transverse filter This equation is solved by simultaneously solving each of its standard Weiner-Hopf matrix equation components Each component is associated with a pair of D subscales, 1 and j As with the standard equation, each component equation is solved iteratively by descending an M,+l dimensional error performance surface that is shaped as a bowl, where M, is the number of samples in D subscale 1 It has M, degrees of freedom represented by the filter's tap weights The optimal solution for each equation is at the bottom of each bowl, as shown in Figure 24 Again, as with the standard case, at each step, for each set of tap weights in the D filter bank

This recursively determined solution is realized by dynamically changing tap weights in each adaptive filter in the D filter bank

3.3.3 Discussion of Generalized Weiner-Hopf Equation

The standard equation results when the special case of uniform sampling is applied. For then, the D Scale has only one subscale and the matrix of vector of vectors collapses to the standard form. Indeed, the underlying model then reverts back from that shown in Figure 23 to the standard model shown in Figure 21.

Another point worth noting is that the generalized equation has a structure that is reminiscent of the fractional adaptive equalizer. But even as the syntax describing both systems is similar, the semantics of the two differ considerably. One difference is that the fractional equalizer uses a higher data rate whereas the D Filter is designed around lower rate D Scale signals.

The interpretation of the solution of this equation is a generalization of the interpretation of the usual equation. The search for a solution of this equation can be inteφreted as the simultaneous descent down multiple bowls of possibly significant lower dimensions. This is illustrated in Figure 21. Each extrema of each bowl corresponds to a vector of possibly complex tap weights of a filter in the D filter bank associated with a D subscale. The full solution of the generalized Weiner-Hopf equation then consists of the vector of vectors, one for each of the D subscales.

3.3 4 Generalized Weiner-Hopf Equation for Nonlinear Systems

Nonlinearities of the type described in Section XXX can be readily accommodated by the generalized Weiner-Hopf equation For the nonlinearities in a channel add additional summations of terms to the impulse response characterization And these added summations manifest themselves by additional elemental, conventiaonal, Weiner-Hopf equations that must be solved simultaneously. The reasoning is identical to that which yielded the original generalized Weiner-Hopf equation for linear systems, as discussed previously

The inteφretation of the solution for a nonlinear system is of an an additional number of bowls, but each a higher dimension, determined by the extent of the nonlinearities

3.4 Benefits of Generalized Weiner-Hopf Based Adaptive Filters

The generalized Weiner-Hopf equation underlies the enhanced design of adaptive filters. Its recursive solution is realized by a set of adaptive filters whose tap weights are varied according to the D Scale non-uniformly sampled inputs. Adaptive filters that are based on D filter design have the following advantages over comparable implementations that rely the standard model. a. Reduced number of taps because each D subscale has far fewer weights than the uniformly sampled case. b.More sensitive to variations in input signal and able to track the variations and volatility. This falls out of the more accurate characterization of a channel. c.Non-calculating or pre-wired D Arithmetic to avoid the deleterious effects of round-off errors in real-time filter calculations, d. Possibly reduced number of iterations for convergence or possibly the time to convergence. This is because the descent down a deep high of order N dimensional bowl is replaced by simultaneous descents down several lower dimensional order and shallow bowls. e. Readily accommodates certain nonlinearities without linearizing away the problem or approximating it. fit enables IIR solutions due to enhanced stability. g.It enables the use of multiple sets of transversal filters, each with potentially far fewer number of taps than the conventional filter. Fewer taps along the end-to-end path of a sample's traversal is generally advantageous.

3.5 Computing Autocorrelation and Cross-Correlation

As with the industry standard digital filter, practical implementation of digital filters requires the accurate computation of the input signal's autocorrelation as well as of the cross correlation of the input and output signals. The same D Scale decomposition technique used to derive an exact numeric expression for the Fourier (or other kernel) transform can also provide accurate numeric calculations of autocorrelation and cross-correlation functions of possibly complex valued signals Further, the derivation is almost identical to that presented previously to compute the convolution

Consider the continuous autocorrelation integral for an input signal

+ oo Ruu(t) = j u(t) u(t+τ) dτ

- oo

Let the underlying D Scale used to sample a signal be comprised of N subscales {pi, .... PN} .

Let δ,_j denote the distance of point P_j in D subscale I to the next D Scale point

Recall that the sampling is uniform at a rate of 1/ p, for each subscale l. Therefore, for points on subscale I, t = n/p, where n= 1 , p, - 1

Further, the correlation integral can be decomposed using Riemann summation over the D subscales. The continuous correlation integral is computed by determining its value separately at points on each D subscales

On each subscale p_j the integration variable τ extends over uniform intervals 1/p_j. Therefore on each subscale τ = k/p_j

Also, on each subscale, the non-uniform interval associated with each uniformly distributed D Scale point leads to dτ -» δ_{j k}

Finally, after substituting these mappings into the Riemann summation that computes the original integral, the following equation emerges Computation of Autocorrelation using the D Scale

Ruu(n/p,) = Σ Σ u(n/p,) u(n/p, + k/p_j) δμ j e {pi, .... p_N} k=0 0<= i < N

Note that, as in the computations of the Fourier transform and convolution, the final result is generated as by concatenating the partial results for each D subscale.

Similarly, consider the continuous cross-correlation integral for input-output signals:

+ 00

Ruy(t) = J u(t) y(t+τ) dτ

- oo

Computation of Cross-correlation using the D Scale

k= _Pj -l Ruy(n p,) = Σ Σ u(n p,) y(n/p, + /p_j) δ_j._k j € {pi, .... p_N} k=0

0<= i < N

Note again, that, as in the computations of the Fourier transform and convolution, the final result is generated as by concatenating the partial results for each D subscale.

The autocorrelation functions can also be computed via the indirect approach of computing the spectrum of the signal and relying on the Weiner-Khintchine Theorem. The D Scale based decomposition of the spectrum would then be Applied.. 3.6 Generalized Weiner (Matched) Filter

3.6.1 Conventional Weiner Filter

The Weiner filter is also based on the Weiner-Hopf equation although it is typically presented as a matched filter. It relies on a-priori knowledge of to correlate the combined input signal and noise against the expected signal. This why it is often referred to as a matched filter, for it aims to match the input to a specific signal. Figure 23 shows the standard continuous Weiner filter. Figure 24 shows the usual discrete realization of the continuous formulation. There, one correlator is applied to uniformly sampled inputs to compute:

N-l r_uy(l) = 1/N Σ u(n) y(n + l) n=0 where 1/N is the sample rate.

3.6.2 Generalized Weiner Filter

The generalized Weiner filter design uses a D Scale data sampler followed by a router to a bank of industry standard Weiner matched filters. The signal to be matched at each Weiner filter in the bank is sampled at the rate of its corresponding D subscale. This is illustrated in Figure 25.

Note that the design is a subset of the most general D filter design architecture shown in Figure 6. Note also from that figure, that there are N vectors of weights. Each weight corresponds to a coefficient in the exact autocorrelation expansion of the continuous autocorrelation function of the input signal, however transient or non-stationary.

3.7 Generalized Kalman Filter

The Kalman filter is also based on the Weiner-Hopf equation But it is most often formulated in the state variable formalism that make the underlying foundation less obvious. The standard Kalman filter is derived from a pair of process (or state) and measurment equations A signal flow graph representation of a linear discrete-time dynamic system is shown in Figure 26 That figure illustrates the following equations

Process (or state) Equation- x(n+l) = F(n+l,n) * x(n) + vι(n) x(n) ^ Unknown state vector of dimension M at time n unit intervals.

F(n+l,n) ^■> Known MxM state transition matrix relating the state of the system at times n+1 and n unit intervals vι(n) -^ Mxl vector representing process noise at time n unit intervals

Measurement Equation: y(n) - C(n) * x(n) + v₂(n)

C(n) ^■■> Known NxM measurement matrix. v (n) "^ Nxl vector representing measurement noise at n unit intervals e^" _k = x(n) - x^"(n) *^ Estimation error between state vector and prior estimate or best estimate about the process prior to time n

P_k = E[e^" _k e^" _k ^τ] "^ Predicted state error covaπance or correlation matrix

The Kalman filteπng problem is to use the entire observed data set, consisting of vectors y(l), y(2),..y(n), to find for each n > 1 , the minimum mean-square estimates of the components of the state x(ι) The generalized Kalman filter is derived using the same analytical steps to derive the standard Kalman filter. The derivation is based on the generalized signal flow graph representation shown in Figure 27. That figure illustrates the following equations:

Process (or state) Equation: xj(l) = F_u(l,k) * x,(k) + v_H(k)

Xι(k) "^ Unknown state vector of dimension M from D subscale i at time k intervals on that D subscale.

X_j(l) ^■> Unknown state vector of dimension M from D subscale j at time 1 intervals on that D subscale. This is the very next point on the D Scale following xι(k).

Fi_j(l,k) Known MxM state transition matrix relating the state of the system at adjacent points on the D Scale, that at 1 intervals on D subscale j and the previous point at interval k on D subscale k.

V_ϋ(k) "^ Mxl vector representing process noise at n unit intervals.

Measurement Equation: y,(n) = Cι(n) * x,(n) + v_2i(n)

Ci(k) "^ Known NxM measurement matrix at a point on the D Scale that is at the k-th interval on D subscale i. v₂i(k) "^ Nxl vector representing measurement noise at a point on the D Scale that is at the k-th interval on D subscale i. e _k,i ⁼ Xi( ) - x^' i(k) "^ Estimation error between state vector and prior estimate or best estimate about the process prior to this D Scale point.

P_k,i = E[e^" _k>i e^" _k ] • Predicted state error covariance or correlation matrix.

The generalized Kalman filtering problem is to use the entire observed data set, consisting of vectors at times on the D Scale yι( l ), ... y-.(\), to find for each time along the D Scale (i.e., each partial time series), the minimum mean-square estimates of the components of the state xι(k) on each D subscale. The only difference in the analysis is that each new sample is mapped to a D Scale point that is associated with a unique D subscale The correlation matrix element in the super-matrix of the generalized Weiner-Hopf equation is identified from its D subscale. The correlation matrix element corresponding to that D subscale must be used for that cycle in the Kalman filter's recursive algorithm. Therefore the only addition to the Kalman iterative procedure is to identify which correlation matrix in the super-matrix to use on each iteration.

The resulting modification to the standard Kalman Filter is shown in the flow chart in Figure 28, with the new steps highlighted Again, the extra step is to fetch the corresponding correlation matrix element for each new sample point.

Note again that the usual Kalman filter algorithm appears for the special case of uniform sampling. Also note that nonlinearities are transparently factored into the calculations because their effects are already embedded into the off-diagonal weights and their corresponding correlation matrix elements.

Claims

1. An N dimensional signal processor comprising a multiresolution sampler outputting a stream of non-uniformly spaced samples resulting in an overall average sampling rate sufficient to satisfy Nyquist criteria to reconstruct the signal, and a digital filter bank where each filter in the bank does filtering of a subset of the samples at different uniform rates such that the stream is uniformly spaced.

2. The N dimensional signal processor of claim 1 , wherein streams of data enter the filter bank at low resolution.

3. The N dimensional signal processor of claim 1 , wherein the non-uniformly spaced samples are constructed according to a D scale.

4. The N dimensional signal processor of claim 2, wherein the non-uniformly spaced samples are constructed according to a D scale.

5. The N dimensional signal processor of claim 1 , wherein at least one filter in the bank receives samples from said sampler based upon a D scale.

6. An N dimensional signal processor having a plurality of taps, each of said plurality of taps receiving data sampled as a subscale rate less than the overall sampling rate.

7. The N dimensional signal processor of claim 6, wherein each subscale is constructed according to a D scale.

8. An N dimensional signal processor comprising banks of filters, each of said filters comprising a plurality of taps, each tap having a low sampling rate, said signal processor having a high sampling rate substantially equal to the sum of the low sampling rates of the plurality of taps.

9. The N dimensional signal processor of claim 8, wherein each of said low sampling rate taps samples data based upon a D scale.

10. A signal processor having an infinite impulse response comprising a feedback loop and a plurality of taps sampling data based on a D scale.

11. A signal processor having a finite impulse response comprising a plurality of taps sampling data based on a D scale.

12. An N dimensional digital filter that extracts information from non-uniform sampling of an N dimensional signal volatility. ?

13. A digital filter bank where each filter in the bank filters samples incoming at

13. A digital filter bank where each filter in the bank filters samples incoming at different rates less than the Nyquist rate resulting in overall average sample rate sufficient to satisfy the Nyquist criteria to reconstruct the signal.

14. A digital filter that extracts information by simultaneous pseudo-random non- uniform sampling.

15. The digital filter of claim 14, wherein the pseudo-random non-uniform sampling employs a D-scale.

16. A digital filter for sharpening spectrum cutoff and minimizing ripples in a frequency spectrum while essentially preserving phase, comprising ...

17. A digital filter for tracking changes from a reference condition comprising ...

18. A digital filter comprising a non-uniform scale multi-resolution input sampler outputting a stream of non- uniformly spaced samples resulting in an overall average sampling rate sufficient to satisfy the Nyquist criteria to reconstruct the signal, a filter bank router a digital filter bank a consolidator arithmetic mappings a sampler controller a phase shifter to modify the phase of output samples for each non-uniform subscale.

19. The digital filter of claim 18 wherein said filter bank router maps each point in the output stream to a transversal filter in a bank of filters associated with a particular non- uniform subscale.

20. The digital filter of claim 18, wherein said digital filter bank implements a vector of tap weights.

21. The digital filter of claim 20, wherein the implementation of a vector of tap weights uses standard FIR and IIR architectures.

22. The digital filter of claim 20, wherein the digital filter is associated with a non- uniform subscale.

23. The digital filter of claim 21 , wherein the digital filter is associated with a non- uniform subscale.

24. The digital filter of claim 18, wherein said digital filter bank associates each filter with a subscale of the non-uniform scale used for sampling.

25. The digital filter of claim 24, wherein a plurality of filters independently sample the incoming data at the rate of its associated non-uniform subscale.

26. The digital filter of claim 24, wherein each of the filters independently sample the incoming data at the rate of its associated non-uniform subscale.

27. The digital filter of claim 18, wherein the consolidator consolidates or interleaves uniformly sampled non-uniform subscale streams into one non-uniformly spaced stream.

28. The digital filter of claim 18, wherein said arithmetic mappings comprise networked tables.

29. The digital filter of claim 18, wherein said arithmetic mappings implement real time calculations on the sampled values.

30. The digital filter of claim 29 wherein said real time calculations comprise ...

31. The digital filter of claim 18, wherein said phase shifter modifies the phase of output samples.

32. The digital filter of claim 31 , wherein said phase shifter modifies the phase of output samples for each non-uniform subscale.

33. Th^digital filter of claim 31 , wherein said phase shifter modulates or controls one or more of positive, negative and phase lock feedback loops.

34. A digital filter having a static design derived from non-uniform scale components

35. A digital filter having and adaptive design derived from non-uniform scale components

36. A method of using a non-uniform subscale filter to modulate input streams based on the output comprising using control signals based on the output, applying the control signals to input samplers.

37. The method of claim 36, wherein the control signals suppress or trigger non- uniform samplers in frequency ranges.

38. The method of claim 36, wherein the sample controller module suppresses or triggers non-uniform samplers in frequency ranges.

39. The digital filter of claim 18, wherein the phase shifter modifies the phase of an output or an input signal.

40. A digital filter compπsing a non-uniform sampler, said sampler comprising A/D converters, wherein each A D converter samples uniformly at a rate determined by the non-uniform sampler's subscale resolution.

41. The digital filter of claim 40, further comprising a dispatcher to fan out the origina analog data stream into multiple fees for multiple non-uniform subscale samplers, wherein each sampler filters out all but its subscale points.